Multi-Layer Backpropagating Neural Network with Generalized Delta Learning Rule

Implementation of a multi-layer back propagating neural network algorithm using the generalized delta rule. Assumes a simple fully connected feedforward network with 1 hidden layer.

Setup

Before running the network, ensure your parameters and data sets are in place. Place the following files in the main directory:

• param.txt
• in.txt
• teach.txt

param.txt should contain 6 lines, each with a single value. The first 3 lines respectively specify, in integers, the number of input, hidden, and output units. The next 3 lines respectively specify, as real values, the learning constant, the momentum constant, and the error criterion.

in.txt should contain the input patterns, with one pattern per line. Each pattern should be a sequence of values, separated by single spaces.

teach.txt should contain the teaching patterns. Each pattern should be a sequence of values, separated by single spaces.

By default, the paramaters and data sets used are for the iris data set.

Run

To run the neural network, navigate to the main directory and run the following command: python main.py

Alternatively, run: make (will clean up *.pyc files before running the main program)

This will launch the text-based user interface with the following options:

	1 - Train network with current settings.
	2 - Give the network a pattern and see the predicted output.
	3 – View network settings.
	4 - Change network settings.
	5 – Exit.

Enter the number for the option you want. Note: The network cannot predict output without being trained first.

File Descriptions

Necessary files for running the main interface:

• main.py contains the code for the text-based user interface

• backprop.py contains the code for the backpropagating neural network algorithm implementation

• neuron.py contains the code for the implementation of individual neurons

• parse.py contains the code for parsing text files containing network parameters, input patterns, and target values

• param.txt see Setup

• in.txt see Setup

• teach.txt see Setup

Helpful but not strictly necessary:

• test.py contains the code for testing various parameter values or data sets
• Makefile contains Make rules for convenient cleanup, running, and testing of the network
• data/* directory containing sample data sets, necessary for running tests in test.py